CLaRa-7B-Instruct is a next-generation Large Language Model architecture released by Apple in 2025, fine-tuned on Mistral-7B. It introduces an innovative Continuous Latent Reasoning mechanism.
CLaRa (Continuous Latent Reasoning approach) is not just an LLM; it is a complete Unified Retrieval-Augmented Generation (RAG) Framework.
By utilizing a shared Continuous Latent Space, it achieves end-to-end optimization of both retrieval and generation processes.
The retriever optimizes for similarity, while the generator optimizes for prediction.
Raw long documents consume massive amounts of tokens and VRAM.
Discrete retrieval steps block gradient backpropagation.
How CLaRa compresses long docs into Memory Tokens for reasoning in latent space.
CLaRa does not simply truncate documents. Instead, it uses a LoRA adapter to train a compressor. This compressor converts raw documents into a fixed number of Memory Tokens (supporting 16× and 128× compression rates).
Utilizing a Differentiable Top-K Estimator, CLaRa allows gradients to backpropagate from the generator's loss back to the retrieval module.
| Model Architecture | Compression (CR) | F1 Score |
|---|---|---|
| PISCO | 16x | 58.55% |
| CLaRa Instruct | 16x | 63.90% |
| Mistral-7B RAG | 1x | 64.24% |
from transformers import AutoModel
# Key Point: trust_remote_code=True is required
model = AutoModel.from_pretrained(
"apple/CLaRa-7B-Instruct",
trust_remote_code=True
).to("cuda")
output = model.generate_from_text(
questions=["Which genus is native to Mexico?"],
documents=[["Document content..."]],
max_new_tokens=64
)